Field Evaluation of Text-Dependent Speaker Recognition in an

Field Evaluation of Text-Dependent Speaker Recognition in an
Access Control Application
Harsh Gupta, Ville Hautamäki, Tomi Kinnunen and Pasi Fränti
Speech and Image Processing Unit,
Department of Computer Science,
University of Joensuu, Joensuu, Finland
{gupta, villeh, tkinnu, franti}@cs.joensuu.fi
Abstract
Vector quantization (VQ) is a widely used matching
algorithm for text-independent speaker recognition. In this
paper we study the use of text-dependent speaker recognition
in practical access control application. We compared dynamic
time warping (DTW) to VQ-based matching using textdependent pass phrases. Our goal was to find out, how fixed
phrase affects speaker recognition performance. We collected
corpus of 21 speakers at the location of access control system
and experimented with two different text-dependent
scenarios: with speaker dependent phrases and with speaker
independent phrases. In both cases, DTW outperforms VQ
matching, or works similar. Also text-independent test were
carried out.
1. Introduction
One of the advantages of speech as a biometric identifier is
user convenience. Firstly, it is possible to give the identity
claim using the same modality as the biometric sample itself
[1]. Secondly, it has been proposed that in a fixed phrase
system, the user could select himself/herself the pass phrase
[2]. Having this in mind, one possible application of speaker
recognition technology is access control into physical
facilities where user convenience is to be preferred over
security. For example in the hospital environment, doctors
and nurses should have a priority access to elevators with
minimal user interaction. Speaker recognition technology
could be used in this case as a biometric access control
method. Furthermore, security of the system is not nearly as
important as user convenience of it.
Having these considerations in mind, traditional speaker
verification scenario might not be the best choice for userconvenient applications. Instead, open-set speaker
identification seems a more feasible choice (Fig. 2). In an
open-set identification system, the unknown sample is
compared with all the voice templates of the authorized
persons. If the best-matching speaker’s score exceeds a
threshold, the user is accepted and otherwise rejected. This
threshold is usually common for all users, but there is one
reason for making it user-depended: some person(s) might be
so called ‘sheep’ so that everyone (also non-authorized
persons) is identified as being this person [3]. For this user,
the threshold should be set higher in general.
In this study, we consider an access control system
prototype implemented by our research group and installed in
our lab door, see Fig. 1. In text-independent speaker
recognition corpus simulations, low error rate is easy to
achieve [4, 5, 6]. On the other hand, we have noticed in
Identification
Speaker
Database
Claim
Verification
Decision
Threshold
Figure 1: Open-set speaker identification system .
practice that low error rate is difficult to achieve in everyday
environment with short test phrases. It is clear that when user
convenience is more important than security, system should
be able to verify the user with very a short phrase. Textdependent speaker recognition achieves speaker recognition
from short pass phrases. We study our access control system
in text-dependent usage scenario. In text-dependent scenario,
we compare the effect of using common phrase or speaker
dependent phrases. Also we compare pattern-matching
algorithm (vector quantization vs. dynamic time warping).
The rest of this paper is organized as follows: In Section 2
access control system is described and Section 3 describes in
detail both DTW and VQ approaches. Section 4 presents the
experimental setup and data collection. Section 5 gives
experimental results and Section 6 gives conclusions.
2. Description of the system
The door to enter our lab has an electronic lock, connected to
a PC via a Bus control unit and Door control unit. MFCC and
VQ based speaker recognition software developed by our
research group, which allows enrolling of speakers into the
database and carrying out speaker recognition and
verification, has been installed on the PC. There is a button
connected to the PC near the door, on pressing which the
software starts to record voice via the microphone near the
door, and stops recording on pressing the button again.
(Figure 2) This unknown speech sample is matched against
enrolled speakers in the speaker database. Then identified
speaker label is sent as a claim to the verification system,
which decides, whether to open the door or not. If approved,
the software sends a message to the Door Control Service
running on the PC, which in turn opens the door lock via the
Door Control Unit.
Step 1 Initialization
D(1,1) = d(1,1)m B(1,1), for j = 2,....,M compute D(i,j) = ∞
Step 2 Iteration
for i = 2,..., N {
for j = 1,...,M compute {
D ( i , j ) = min [D ( i − 1 , p ) + d ( p , j ) ]
1≤ p ≤ M
B ( i , j ) = a r g m i n [D ( i − 1 , p ) + d ( p , j ) ]
1≤ p ≤ M
}}
3. Template Matching Methods
Figure 3: The Dynamic Programming Algorithm [8]
We applied both the dynamic time warping and vector
quantization template matching methods. In both methods,
template Ci is first created for each enrolled speaker i and then
unknown sample X is matched against all templates. Matching
outputs a dissimilarity score D(.), from where speaker with
minimum score is selected as the identified speaker.
Id =arg min {D ( X , C i )}
i =i ... N
Mel Frequency Cepstral Coefficients (MFCC) features
were used to represent raw speech signal into compact but
effective representation that is more stable and discriminative
than the original signal [6]. First the pre-emphasized and
windowed speech frame is converted into spectral domain by
the Fast Fourier Transform (FFT). The magnitude spectrum is
then smoothened by a bank of triangular bandpass filters that
emulate the critical band processing of the human ear. Each
of the bandpass filters computes a weighted average of that
subband, which is then compressed by logarithm. The log
compressed filer outputs are then decorrelated using the
Discrete Cosine Transform (DCT). The zeroth cepstral
coefficient is discarded since it depends on the intensity of the
frame.
3.1. Dynamic Time Warping
Dynamic time warping uses the principle of dynamic
programming [principle of optimality], in order to compute
overall distortion between the two speech templates.
Comparing the template with incoming speech might be
achieved via a pairwise comparison of the feature vectors in
each. The problem with this approach is that if constant
window spacing is used, the length of the input and stored
sequences is unlikely to be the same. Moreover, within a
word, there will be variation in the length of individual
phonemes. The matching process needs to compensate for
length differences and take account of the non-linear nature of
the length differences within the words.
The dynamic time warping algorithm achieves this goal,
it finds an optimal alignment between two sequences of
feature vectors, which allows for stretched and compressed
sections of the sequence [7].
We can arrange the two sequences of observations on the
sides of a grid with the unknown sequence on the bottom and
the stored template up on the left. Both sequences start on the
bottom left of the grid. Inside each cell we can place a
distance measure comparing the corresponding elements of the
two sequences.
Fixed Template (Spk1, Male)
Figure 2: Speaker recognition system in access control.
Step 3 Backtracking and Termination
The optimal (minimum) distance is D(N,M) and the
optimal path is (s1, s2,..., sN) where sN = M and si = B(i+1, si+1,
i = N-1, N-2, ..., 1
60
100 Frames = 2 secs
50
40
30
20
Spk1(M), Spk1(M)
Spk1(M), Spk2(F)
Spk1(M), Spk3(M)
10
0
0
20
40
60
Testing Sample
80
Frames
Figure 4: Optimal DTW paths
To find the best match between these two sequences we
can find a path through the grid which minimizes the total
distance between them. Figure 4 shows the optimal paths
obtained by comparing the template of a speaker with his own
testing sample and two other speakers, one male and other
female. The DTW algorithm is designed to exploit some
observations about the likely solution to make the comparison
between sequences more efficient. Instead of finding all
possible routes through the grid that satisfy these constraints,
the DTW algorithm works by keeping track of the cost of the
best path to each point in the grid. The purpose of DTW is to
produce a warping function that minimizes the total distance
between the respective points of these phrases. We used
symmetric DTW with slope constraint = 0. The distances in
DTW were calculated using Euclidean distance metric.
3.2. VQ Matching
In the VQ-based approach to speaker identification, the
speaker models are formed by clustering the speaker's feature
vectors into K disjoint clusters [4, 5, 9]. Each cluster is
represented by a code vector ci, which is the centroid (average
vector) of the cluster. The resulting set of code vectors {c1,..,
ck} is called a codebook, and it serves as the model for the
speaker. However, in these experiments we used raw feature
set extracted from the training speech as a speaker model. This
was done because speech segment was so short that clustering
was not practical.
The matching function in the baseline VQ-based speaker
recognition [9] is the quantization distortion between the two
vector sets to be compared. Given a feature vector xi
generated by the unknown speaker, and a codebook C = {c1
…ck }, the total quantization distortion DQ is given by
DQ ( X , C ) =
1
X
∑ min d (x , C )
xi ∈ X
c j ∈C
Table 1: Identification error rates in speaker identification
Speaker
Speaker
Dependent
Indpendent
DTW
VQ
DTW
VQ
Test -Train
S1 – S2
4.8
0
14.3
19.1
S2 – S1
0
19.1
19.1
28.6
(1)
i
where d(·,·) is a distance metric defined over the feature
space. Typically Euclidean metric is used as the distance
measure. The normalization factor 1/|X | is the same for all
speakers, and therefore it does not change the order of the
speakers in the matching result.
4. Experimental Setup
In order to analyze the effectiveness of both methods in the
real world applications, all the samples were recorded by the
Door Access Control System, in place in our lab at Speech
and Image Processing Unit (SIPU), University of Joensuu.
This allowed ambient noises and room acoustics to interfere
with the recordings. Samples from 21 speakers were
collected, out of which 7 were females and rest males, none
of which were native English speakers. All the phrases were
in English. One of the two training phrases was common to
all speakers. This phrase was used in common phrase testing.
The second phrase was unique for every speaker. In both
cases, the same corresponding sentence was used for testing.
The speaker-dependent phrase varied in length for every
speaker, and no emphasis on the phonetic content of these
phrases was laid. Common phrase was “Every salt breeze
comes from the sea” and speaker dependent phrases were
sentences from newspaper clippings (Appendix A).
Recordings were done at the location of entrance door of our
lab. All samples were recorded at 22 KHz, in 16-bit PCM
WAV format. The recorded samples were manually aligned
by removing the initial and trailing silence.
As can be inferred from the table, DTW gives better or
comparable results with VQ-matching for all the cases.
Moreover, it can be seen that the samples collected when the
user has become acquainted with the system and the fact that
he/she is being recorded doesn’t bother him/her anymore,
serve as better training samples to compute the templates.
Thus training samples should be as long as permissible by the
application where speaker recognition system is being
deployed, so that the length of the sample can compensate the
variation in the speech pattern of the user.
5.1. Sample length
In order to assess the effect of the length of training and
testing samples on speaker identification, we first aligned
both the samples of the same speaker using DTW and then
made 10 equal sections of the optimal alignment path found..
The feature vector files of the samples were copied from the
beginning to the point indicated by the ith section of the
minimal path. Figure 5 summarizes the results obtained. The
‘VQ(TI) Matching’ plot gives the error rates of using VQMatching for Text-Independent scenario.
DTW gives much less error rates on smaller samples, in
both speaker independent and speaker dependent scenarios, as
per our expectations. This makes DTW fit for use in case of
speaker identification using small pass phrases, where it
outperforms VQ by a good amount. If samples are long
enough then both VQ and DTW perform on a similar level,
which is much better than the Text-Independent case.
4.1. Feature Extraction
MFCC features have been used for evaluating the
performance of both DTW and VQ-Matching. We used
MFCC features of order 12, with window size = 30 ms. and
window shift = 20 ms. Features were extracted using the
command line automatic speaker identification software
developed at our lab, hereafter referred to as "sprofiler". First
the features were extracted for all the samples. Then VQmatching based speaker identification results were evaluated
using sprofiler and for DTW based results, code available at
Helsinki University of Technology website was used.
5. Results
We evaluated the error rates on our corpus for both the textdependent scenarios: speaker independent and speaker
dependent. The results have been summarized in Table 1. The
row corresponding to ‘S1 – S2’ quotes the results obtained
when the voice samples obtained in first recording were used
as testing samples and the samples collected in second
recording were used as training samples. The next row
corresponding to ‘S2 – S1’ in the first column quotes the
results by interchanging the training and testing samples that
were used in the previous test.
Figure 5: Effect of sample length on VQ and DTW matching
5.2. Verification
The verification results for are presented in Fig. 6 as a
receiver operating curve (ROC) showing the genuine
acceptance rate as a function of the false acceptance rate. The
equal error rates are also summarized in Table 2.
phrase should be as short as possible, and thus we recommend
to use DTW as the default method. In future, effects of the
pass phrase and intersession variability should be studied. The
results should be also validated on a larger database.
7. Acknowledgements
Work of Harsh Gupta was supported by the National
Technology Agency of Finland (TEKES) (project
“Puhetekniikan
uudet
menetelmät
ja
sovellukset”,
http://cs.joensuu.fi/pages/pums, TEKES dnro 8713103).
8. References
[1] Heck L. and Genoud D., “Combining Speaker and
Speech Recognition Systems”, Proc. Int. Conf. on Spoken
Language Processing, p 1369-1372, 2002.
[2] Rosenberg, Aaron E. / Parthasarathy, S. (1997): "Speaker
identification with user-selected password phrases", Proc. 5th
European Conf. on Speech Communication and Technology, p
1371-1374, 1997
[3] Campbell, J.P. “Speaker Recognition: a Tutorial”,
Proceedings of the IEEE, Vol. 85, no. 9, 1997, p. 1437-1462
[4] Kinnunen, T., Karpov, E. and Fränti, P., “Real-Time
Speaker Identification and Verification”, IEEE Transactions
on Speech and Audio Processing (accepted for publication).
[5] Saastamoinen, J., Karpov, E., Hautamäki, V. and Fränti,
P.,"Accuracy of MFCC based speaker recognition in Series 60
device", EURASIP Journal on Applied Signal Processing,
(accepted for publication).
Figure 6: ROC Curves for different scenarios
Algorithm
DTW
VQ
Common
Phrase
10.2
19.0
Speaker
Dependent Phrase
9.5
7.9
Table 2: Equal Error Rates
For both the common and speaker-dependent phrase,
DTW outperforms VQ. Especially for the common phrase in
which the utterance is much shorter, the absolute difference in
the both false acceptance and false rejection rate is about 5-10
%. For the speaker-dependent case with longer utterances, the
differences are smaller, but DTW outperforms VQ at low
false acceptance rates. On the other hand, in the user convenient end of the tradeoff curve the differences are
smaller. Based on these results, we suggest using DTW if the
pass phrases are short or security (low false acceptance rate)
are important. For longer pass phrases and user-convenience,
the method does not play so important role.
6. Conclusions
In the view of results presented here, the DTW approach
outperforms VQ if the pass phrases are short. For longer pass
phrases the two approaches perform essentially similar.
However, from the user convenience point of view, the pass
[6] Kinnunen, T., Hautamäki, V., Fränti, P., “"Fusion of
Spectral Feature Sets for Accurate Speaker Identification”,
Proc.9th International Conference Speech and Computer
(SPECOM'2004), pp. 361-365, St. Petersburg, Russia,
September 20-22, 2004
[7] Sakoe, H. and Chiba, S. (1978) ‘Dynamic programming
algorithm optimization for spoken word recognition,’ IEEE
Trans. ASSP, vol.26, no. 1,43-49.
[8] Huang X., Acero A., Hon H., (2001) ‘Spoken language
processing: a guide to theory, algorithm, and system
development’, Prentice-Hall, New Jersey, USA
[9] Soong, F.K., Rosenberg, A.E., Juang, B. H. and Rabiner,
L.R., “A vector quantization approach to speaker recognition”,
AT & T Technical Journal, Vol. 66, p. 14-26, 1987.
Appendix A
- Finland's guards are threatening to start a new strike on the week of the Midsummer holiday.- According to a labour market advisor of the
Central Organisation of Finnish Trade Unions, poor treatment of Estonians working in Finland is an everyday occurrence. - A report
published in Sweden on Tuesday on the response of the country’s officials to the aftermath of the Asian tsunami sharply criticises the
response of Swedish officials to the disaster. - The decision by the Finnish Frontier Guard to stop inspecting passports of travellers is not a
violation of the Schengen Treaty. - The change in conditions in the passenger traffic on the Baltic Sea is plain to read from the financial
reports.- Though it is early days yet, the summer of 2005 seems frighteningly similar to the one before. - The Labour Court ruled on
Monday that the tasks the Frontier Guard had been allocating to non-union personnel at crossing points could not be labelled safety work. A fresh study indicates that Finnish elderly people tend to be more open-minded than senior citizens in other European countries. Persistent rain on Saturday made a mess of the time-honoured tradition of new upper secondary graduates. Finnish pulp and paper mills
will remain shut down for the time being. - An extensive Trafficking in Persons Report by the United States Department of State ranks
Finland among countries which do not fully comply - The decision by the Finnish Frontier Guard to stop inspecting passports of travellers
is a violation of the Schengen Treaty after all.- Fortum’s share price has more than doubled in just over a year, and it had been expected
that the state would take advantage of the situation and sell its extra stock. - The Finnish Parliament is to debate the proposed constitution
for the European Union in the autum. - The EU has existed for decades without a constitution, and can certainly continue. - With the
introduction of the euro, consumers' ability to perceive the cost of things has become weaker. - Another interesting point is that the nominal
value of the currency seems to bear significance in relation to consumer price awareness - Employees of the Leaf sweets factory in Turku
walked off their jobs again on Thursday afternoon. - In the future, an estimated 150,000 people will pass through Helsinki's new Kamppi
Center every day. - In one of the stranger stories to come to light this week, two large stray dogs apparently scared passers-by. - Russia has
denied that its military planes had violated Finnish airspace over the Gulf of Finland.
Download PDF